1 00:00:03,000 --> 00:00:05,120 Evan: Welcome to Daily Paper Cast! 2 00:00:05,214 --> 00:00:13,464 Ashley: Today's paper comes from the Hugging Face daily paper list of October 14, 2026, and it's already garnered 40 upvotes. 3 00:00:13,512 --> 00:00:20,492 Evan: The title of the paper is 'Jeva: System One Models from First Principles.' 4 00:00:20,544 --> 00:00:26,144 Ashley: It's authored by Qi Li and Xinchao Wang, with Li Qi as the corresponding author. 5 00:00:26,144 --> 00:00:28,924 They're from the National University of Singapore. 6 00:00:29,048 --> 00:00:34,608 Evan: Ashley, can you start us off with a bit of background on what this paper addresses? 7 00:00:34,656 --> 00:00:35,916 Ashley: Sure, Evan. 8 00:00:35,916 --> 00:00:47,756 System One models are designed for direct and low-latency decision-making, appearing often in scenarios that require immediate choices like identifying documents or selecting tools. 9 00:00:47,808 --> 00:00:57,348 Evan: Right, and the authors argue that recent approaches, which predict answers from serialized text, don't adequately respect the structure of the decision problem. 10 00:00:57,408 --> 00:00:58,538 Ashley: Exactly. 11 00:00:58,538 --> 00:01:03,828 They emphasize four crucial structural requirements that are often underappreciated. 12 00:01:03,828 --> 00:01:11,948 First, models need to be null-aware, meaning they should distinguish between having no answer and uncertainty among the given options. 13 00:01:12,000 --> 00:01:15,140 Evan: And what are the other structural requirements? 14 00:01:15,192 --> 00:01:17,922 Ashley: The second is permutation-invariance. 15 00:01:17,922 --> 00:01:22,252 This means that reordering the options shouldn't change the decision outcome. 16 00:01:22,252 --> 00:01:30,202 Third is option-scalability, which allows for varying the number of candidate options without needing to redefine the output space. 17 00:01:30,202 --> 00:01:38,232 Finally, cacheability, where computation over frequently recurring alternatives should be reusable across different requests. 18 00:01:38,280 --> 00:01:39,590 Evan: That's interesting. 19 00:01:39,590 --> 00:01:45,500 So the paper introduces a new architecture called Jeva that addresses all these requirements, right? 20 00:01:45,552 --> 00:01:46,412 Ashley: Yes. 21 00:01:46,412 --> 00:01:53,712 Jeva is a decision-native architecture that independently represents both the request and each option using a shared encoder. 22 00:01:53,712 --> 00:02:00,212 The real novelty lies in how it refines only the request state while keeping the option memory immutable. 23 00:02:00,324 --> 00:02:05,744 Evan: Can you break down how this refinement works and why it's beneficial? 24 00:02:05,808 --> 00:02:12,008 Ashley: They use a compact decision stack where the request state is read against this immutable option memory. 25 00:02:12,008 --> 00:02:18,698 This enables interactions between the request and the options without altering the original option representations. 26 00:02:18,698 --> 00:02:25,728 It ensures that important properties such as permutation-invariance and option-scalability are maintained throughout. 27 00:02:25,836 --> 00:02:26,646 Evan: Got it. 28 00:02:26,646 --> 00:02:29,496 And how does Jeva handle decision readouts then? 29 00:02:29,544 --> 00:02:38,974 Ashley: Jeva uses symmetric aggregation to yield permutation-equivariant probabilities, meaning the order of options doesn't affect the outcome. 30 00:02:38,974 --> 00:02:54,824 Additionally, it supports variable-size candidate sets through shared scoring mechanisms and includes a coverage estimate to distinguish between answer absence and competition among candidates, which enables explicit null outcomes. 31 00:02:54,888 --> 00:02:56,378 Evan: That sounds powerful. 32 00:02:56,378 --> 00:03:06,448 So, this architecture separates computation tasks, making option representations reusable and maintaining efficient handling of variations in candidate sets. 33 00:03:06,504 --> 00:03:07,494 Ashley: Exactly. 34 00:03:07,494 --> 00:03:16,604 Jeva combines strong typed-decision performance with near-perfect detection of null outcomes and stable choices regardless of how options are ordered. 35 00:03:16,604 --> 00:03:24,664 It also scales effectively to thousands of candidates, significantly reducing latency by reusing computations across requests. 36 00:03:24,720 --> 00:03:29,780 Evan: What practical impacts do the authors foresee for this innovation? 37 00:03:29,832 --> 00:03:35,682 Ashley: They anticipate that Jeva’s architecture will streamline processes that require rapid decision-making. 38 00:03:35,682 --> 00:03:44,452 It’s particularly relevant in conversational AI, tool selection, and document retrieval tasks where decision latency and accuracy are critical. 39 00:03:44,496 --> 00:03:49,406 Evan: This could indeed be a transformative approach for many applications. 40 00:03:49,406 --> 00:03:53,456 And that's the background, objectives, and contributions covered. 41 00:03:53,520 --> 00:03:56,570 Ashley: Yes, that wraps up the Introduction section. 42 00:03:56,570 --> 00:03:59,920 Let's move on to the methods proposed in this paper next. 43 00:04:01,178 --> 00:04:05,008 Evan: Alright, let's dive into the methods behind Jeva. 44 00:04:05,008 --> 00:04:11,538 Ashley, can you kick us off with how Jeva independently encodes the request and each option? 45 00:04:11,594 --> 00:04:12,784 Ashley: Of course. 46 00:04:12,784 --> 00:04:18,384 Jeva uses a shared language encoder to separately process the request and each option. 47 00:04:18,384 --> 00:04:24,764 The request includes the state information, the question or decision context, and the type of decision. 48 00:04:24,764 --> 00:04:34,394 Each option is encoded without including the current request context or other options, ensuring that each one is contextualized independently. 49 00:04:34,442 --> 00:04:38,862 Evan: So, how does this separate encoding benefit the method? 50 00:04:38,906 --> 00:04:44,776 Ashley: This separation allows Jeva to reuse the encoded representations of options. 51 00:04:44,776 --> 00:04:55,326 For example, if the same option appears in different requests, the encoding for that option does not need to be recomputed, which saves time and computation. 52 00:04:55,370 --> 00:04:56,730 Evan: That makes sense. 53 00:04:56,730 --> 00:05:02,450 And once these encodings are in place, how does Jeva handle the actual decision-making? 54 00:05:02,498 --> 00:05:14,308 Ashley: After the initial encoding, Jeva refines only the request state by comparing it against a memory that holds the immutable option vectors. 55 00:05:14,308 --> 00:05:24,228 This refinement happens through a sequence of decision blocks that read the request state and aggregate evidence from the option memory. 56 00:05:24,228 --> 00:05:33,238 The aggregation is permutation-invariant, meaning that reordering the options doesn't affect the decision outcomes. 57 00:05:33,290 --> 00:05:34,380 Evan: Interesting. 58 00:05:34,380 --> 00:05:41,790 So, how exactly does this refinement process ensure permutation-invariance and efficient scalability? 59 00:05:41,834 --> 00:05:48,694 Ashley: Jeva uses symmetric aggregation within its decision blocks to maintain permutation-invariance. 60 00:05:48,694 --> 00:05:56,074 This symmetric aggregation reads and integrates evidence from all options in such a way that the order doesn't matter. 61 00:05:56,074 --> 00:06:01,334 This ensures that the decision-making process is robust to the ordering of candidates. 62 00:06:01,394 --> 00:06:05,734 Evan: Can you shed some light on how these decision blocks operate? 63 00:06:05,786 --> 00:06:06,746 Ashley: Sure. 64 00:06:06,746 --> 00:06:17,356 Each decision block in Jeva uses a multi-head attention mechanism where the query is derived from the request state, and the keys and values come from the option memory. 65 00:06:17,356 --> 00:06:23,606 The attention mechanism computes weights for each option that indicate their relevance to the request state. 66 00:06:23,606 --> 00:06:33,886 The weighted option vectors are then aggregated to refine the state, which is subsequently normalized and passed through a feed-forward network for further refinement. 67 00:06:33,998 --> 00:06:39,198 Evan: And how does Jeva handle scalability in terms of the number of options? 68 00:06:39,242 --> 00:06:48,462 Ashley: Jeva supports scalability through shared scoring mechanisms that accommodate varying numbers of options without expanding the output parameters. 69 00:06:48,462 --> 00:06:55,782 It also includes a caching mechanism where the representations of options can be reused across different requests. 70 00:06:55,782 --> 00:07:05,142 This is particularly useful for large-scale applications where the same options might be evaluated repeatedly across different decision contexts. 71 00:07:05,186 --> 00:07:15,166 Evan: What about the architecture's ability to distinguish between having no valid option and uncertainty among closely competing options? 72 00:07:15,166 --> 00:07:16,786 How is that achieved? 73 00:07:16,880 --> 00:07:18,590 Ashley: Great question, Evan. 74 00:07:18,590 --> 00:07:25,540 Jeva includes a coverage estimate mechanism that calculates the probability of having at least one acceptable option. 75 00:07:25,540 --> 00:07:36,370 This probability is combined with the individual option probabilities to produce a final distribution that includes the likelihood of null or 'no suitable option' outcomes. 76 00:07:36,370 --> 00:07:42,250 This allows the model to explicitly signal when none of the options fit the decision context. 77 00:07:42,314 --> 00:07:49,014 Evan: And how effective is Jeva in practical terms of latency and computational efficiency? 78 00:07:49,058 --> 00:07:56,118 Ashley: Jeva significantly reduces latency by reusing option representations and leveraging efficient encodings. 79 00:07:56,118 --> 00:08:07,078 For example, it can scale to thousands of candidates and still achieve a median latency speedup of over 20 times compared to conventional methods, thanks to cross-request reuse. 80 00:08:07,078 --> 00:08:15,758 This means that even when processing large candidate sets, Jeva can make rapid decisions without reprocessing all the options from scratch. 81 00:08:15,818 --> 00:08:17,188 Evan: Impressive. 82 00:08:17,188 --> 00:08:24,638 So, beyond the core architecture, what kind of evaluations did they use to test Jeva's performance? 83 00:08:24,698 --> 00:08:28,218 Ashley: The paper reports a variety of empirical evaluations. 84 00:08:28,218 --> 00:08:36,258 They tested Jeva on different datasets including BANKING77 and MASSIVE. 85 00:08:36,258 --> 00:08:45,698 The model showed strong performance in typed decision tasks with nearly perfect null detection and consistent decisions across different permutations of the candidates. 86 00:08:45,806 --> 00:08:57,186 Evan: So, in summary, Jeva's method combines independent encoding, efficient state refinement, and a robust decision module to ensure scalability and efficiency. 87 00:08:57,186 --> 00:09:06,406 It addresses key structural requirements like permutation-invariance, null-awareness, and option scalability using innovative mechanisms. 88 00:09:06,458 --> 00:09:07,608 Ashley: Exactly. 89 00:09:07,608 --> 00:09:10,298 And that covers the methods section of the paper. 90 00:09:10,298 --> 00:09:15,498 Up next, we'll look into the experiments and results that validate Jeva's performance. 91 00:09:16,755 --> 00:09:23,855 Evan: Now that we've covered the methods, let's dive into the experiments and results that showcase Jeva's capabilities. 92 00:09:23,855 --> 00:09:28,515 Ashley, can you start us off with an overview of how they tested Jeva? 93 00:09:28,563 --> 00:09:29,733 Ashley: Sure, Evan. 94 00:09:29,733 --> 00:09:34,653 The authors conducted a series of empirical evaluations to measure Jeva's performance. 95 00:09:34,653 --> 00:09:44,523 They used a common suite of benchmarks, including BANKING77, MASSIVE, and three slices of Mind2Web. 96 00:09:44,523 --> 00:09:52,843 Additionally, they assessed Jeva against Typed Decisions—a dataset from LocalLLaMA, consisting of 2,000 test decisions. 97 00:09:52,899 --> 00:09:54,929 Evan: That sounds comprehensive. 98 00:09:54,929 --> 00:10:00,119 What specific metrics and scenarios did they focus on in these evaluations? 99 00:10:00,171 --> 00:10:12,191 Ashley: They focused on several key metrics, such as first-argmax accuracy, which measures the correctness of the top-ranked option, null-detection AUROC, and permutation stability. 100 00:10:12,191 --> 00:10:22,671 For example, in BANKING77 and MASSIVE, they tested both 8 and 16 candidate configurations. 101 00:10:22,671 --> 00:10:28,571 They also looked at how well Jeva performs under different permutations of the candidate sets. 102 00:10:28,635 --> 00:10:34,735 Evan: How did Jeva compare to other models in terms of first-argmax accuracy? 103 00:10:34,779 --> 00:10:39,399 Ashley: On first-argmax accuracy, Jeva excelled across multiple benchmarks. 104 00:10:39,399 --> 00:10:50,379 It outperformed other models on seven out of eight slices in full-coverage accuracy, achieving over 90% on BANKING77 and MASSIVE. 105 00:10:50,379 --> 00:11:05,199 On Mind2Web cross-task, cross-website, and cross-domain selection, Jeva showed strong performance with scores of 79.99%, 62.25%, and 66.23%, respectively. 106 00:11:05,319 --> 00:11:09,139 Evan: What about Jeva's performance in null detection? 107 00:11:09,195 --> 00:11:12,435 Ashley: Jeva's null-detection capabilities were impressive. 108 00:11:12,435 --> 00:11:25,585 For instance, in BANKING77 with 16 candidates, Jeva achieved a 98.66% AUROC for null detection. 109 00:11:25,585 --> 00:11:34,725 It could correctly identify 93.88% of answer-absent cases at a 5.02% false-positive rate. 110 00:11:34,725 --> 00:11:44,575 This demonstrates Jeva's ability to effectively distinguish between instances where no valid option is present and those where there is uncertainty among options. 111 00:11:44,619 --> 00:11:49,739 Evan: How did Jeva perform in terms of permutation stability? 112 00:11:49,803 --> 00:12:06,243 Ashley: Jeva exhibited zero decision flips across 600 Choice examples under various permutations, with a mean total variation of 3.17e-8 and an all-order accuracy of 76.67%. 113 00:12:06,243 --> 00:12:17,403 This indicates that Jeva's decisions remain consistent regardless of the order in which candidate options are presented, showcasing the permutation-invariance of its architecture. 114 00:12:17,451 --> 00:12:23,371 Evan: What about latency and scalability, particularly with large candidate sets? 115 00:12:23,427 --> 00:12:28,667 Ashley: For scalability, Jeva supports up to 2,048 candidates efficiently. 116 00:12:28,667 --> 00:12:40,077 In ToolRet tests with distinct query cohorts ranging from 32 to 2,048 candidates, Jeva maintained an aggregate accuracy of 78.40%. 117 00:12:40,077 --> 00:13:03,647 It achieved substantial latency reduction as well, with batching techniques that brought down the p50 latency from 31.730 seconds to 7.519 seconds at 1,024 candidates and from 94.459 to 12.040 seconds at 2,048 candidates—a speedup of 7.85 times. 118 00:13:03,699 --> 00:13:10,119 Evan: How did they evaluate Jeva's ability to reuse computations across different requests? 119 00:13:10,209 --> 00:13:14,639 Ashley: They tested Jeva's exact-set and hybrid caching capabilities. 120 00:13:14,639 --> 00:13:34,429 In matched MLDR queries with novel-set priming, hybrid caching achieved speedups of up to 27.98 times in p50 latency at 100% option reuse, bringing p50 latency down from 2,102.0 milliseconds to 75.1 milliseconds. 121 00:13:34,429 --> 00:13:46,739 In shared-bank scenarios for Typed Decisions, caching avoided encoding 99.2% of candidate rows, resulting in significant efficiency gains without changing the outcomes. 122 00:13:46,803 --> 00:13:49,313 Evan: That's impressive efficiency. 123 00:13:49,313 --> 00:14:05,563 So essentially, Jeva has validated its structural improvements through rigorous testing and demonstrated substantial benefits in terms of accuracy, null detection, permutation stability, and latency reduction. 124 00:14:05,619 --> 00:14:07,259 Ashley: Precisely, Evan. 125 00:14:07,259 --> 00:14:19,579 The experiments clearly demonstrate that Jeva's innovative architecture and mechanisms enable it to perform exceptionally well across various benchmarks and practical scenarios. 126 00:14:19,635 --> 00:14:23,085 Evan: That completes the Experiments and Results section. 127 00:14:23,085 --> 00:14:29,755 It's fascinating to see such detailed evaluations backing up the proposed benefits of Jeva. 128 00:14:29,811 --> 00:14:39,051 Ashley: Now that we've covered the experiments, we'll move on to discuss related work and contextualize Jeva's contributions within the wider research landscape. 129 00:14:40,331 --> 00:14:50,221 Evan: Now that we've delved into the experiments and results demonstrating Jeva's performance, it's time to place this work within the context of existing research. 130 00:14:50,221 --> 00:14:54,061 Ashley, can you walk us through the Related Work section? 131 00:14:54,155 --> 00:14:55,435 Ashley: Evan. 132 00:14:55,435 --> 00:15:12,505 The authors organize their Related Work section into four main categories: direct and typed decision models, independent encoding and late interaction, set functions and latent attention, and null awareness and calibration. 133 00:15:12,597 --> 00:15:16,717 Evan: Let's start with direct and typed decision models. 134 00:15:16,717 --> 00:15:18,217 What does this involve? 135 00:15:18,299 --> 00:15:26,109 Ashley: Direct and typed decision models aim to produce structured decisions without requiring sequential generated outputs. 136 00:15:26,109 --> 00:15:34,909 These models typically use various readout mechanisms to derive decisions directly from the representations of requests and candidates. 137 00:15:34,973 --> 00:15:39,173 Evan: Can you give some examples of these readout mechanisms? 138 00:15:39,221 --> 00:15:40,051 Ashley: Sure. 139 00:15:40,051 --> 00:15:53,641 Some models rely on answer-token logits or task-specific prediction heads, like the SemIf and JevK5 models, which extract option probabilities from these logits without using autoregressive decoding. 140 00:15:53,641 --> 00:16:01,941 Intern-Decision, for instance, reads distributions over permitted answer symbols using decision markers embedded in the outputs. 141 00:16:02,057 --> 00:16:03,097 Evan: I see. 142 00:16:03,097 --> 00:16:06,977 And how does Jeva differ from these existing approaches? 143 00:16:07,037 --> 00:16:17,127 Ashley: Jeva separates the reusable semantic encoding from the set-dependent judgment stage, which allows the independent text representation of requests and options. 144 00:16:17,127 --> 00:16:29,117 This enables request-independent reuse of candidate content and avoids treating the decision set as an ordered sequence, which can affect predictions based on incidental candidate positions. 145 00:16:29,165 --> 00:16:31,365 Evan: That's a key distinction. 146 00:16:31,365 --> 00:16:36,665 Moving on, what about independent encoding and late interaction? 147 00:16:36,725 --> 00:16:47,655 Ashley: Independent encoding and late interaction refer to methodologies in which queries and candidates are processed separately, allowing for efficient and reusable representations. 148 00:16:47,655 --> 00:16:57,425 A prominent example is Dense Passage Retrieval, where questions and passages are encoded independently, and corpus representations are precomputed. 149 00:16:57,485 --> 00:17:02,645 Evan: How does this concept apply specifically in the context of Jeva? 150 00:17:02,693 --> 00:17:10,853 Ashley: In Jeva, the request state is refined against an immutable option memory, and the model leverages reusable representations. 151 00:17:10,853 --> 00:17:22,193 This approach allows the model to adapt the request state based on the candidate set while maintaining fixed option representations, which cannot be achieved by independent scoring mechanisms. 152 00:17:22,253 --> 00:17:23,833 Evan: Interesting. 153 00:17:23,833 --> 00:17:28,463 Let's discuss the set functions and latent attention. 154 00:17:28,463 --> 00:17:30,333 What does this entail? 155 00:17:30,389 --> 00:17:38,049 Ashley: Set functions and latent attention frameworks are focused on ensuring permutation invariance in set modeling. 156 00:17:38,049 --> 00:17:45,409 For example, Deep Sets establish a basis for this through shared transformations and symmetric aggregation. 157 00:17:45,409 --> 00:17:53,409 The Set Transformer extends this by enabling attention-based interactions among elements, offering fine-grained control. 158 00:17:53,513 --> 00:17:57,333 Evan: And how does Jeva leverage these concepts? 159 00:17:57,389 --> 00:18:07,289 Ashley: Jeva integrates symmetric aggregation and shared scoring within its decision blocks, ensuring permutation-equivariant probabilities. 160 00:18:07,289 --> 00:18:16,809 By refining the request state using the option memory, Jeva maintains the advantages of set-dependent judgment and reusability. 161 00:18:16,809 --> 00:18:25,329 This means it can adapt to variable-size candidate sets without changing parameter shapes, which is crucial for scalability. 162 00:18:25,433 --> 00:18:26,053 Evan: Got it. 163 00:18:26,053 --> 00:18:30,173 Finally, what about null awareness and calibration? 164 00:18:30,221 --> 00:18:40,981 Ashley: Null awareness and calibration involve recognizing when none of the candidates are suitable options, a key aspect in many decision-making scenarios. 165 00:18:40,981 --> 00:18:52,641 This has been studied, for instance, in SQuAD 2.0 for answerability and in selective classification which allows abstention to manage prediction risk. 166 00:18:52,785 --> 00:18:57,045 Evan: How does Jeva address null awareness in its design? 167 00:18:57,101 --> 00:19:06,271 Ashley: Jeva introduces a coverage gate that distinguishes between the absence of a valid candidate and uncertainty among competing candidates. 168 00:19:06,271 --> 00:19:19,521 This enables an explicit null outcome when no acceptable option is present, allowing the model to handle cases where it’s better to indicate 'none of the above' rather than making an unsuitable choice. 169 00:19:19,521 --> 00:19:25,941 This is complemented by post-training calibration techniques to ensure probability estimates are reliable. 170 00:19:25,997 --> 00:19:42,617 Evan: So, Jeva builds on existing research and integrates advances in decision models, independent encoding, set functions, and null awareness to present a comprehensive decision-making architecture. 171 00:19:42,677 --> 00:19:44,167 Ashley: Exactly, Evan. 172 00:19:44,167 --> 00:19:49,997 That covers the Related Work that contextualizes Jeva’s contributions within the broader field. 173 00:19:51,296 --> 00:19:58,866 Evan: As we wrap up, let's recap the key contributions and takeaways from this fascinating paper on Jeva. 174 00:19:58,956 --> 00:19:59,716 Ashley: Evan. 175 00:19:59,716 --> 00:20:11,106 Jeva introduces a novel approach to decision-making models by addressing four critical requirements: null-awareness, permutation-invariance, option-scalability, and cacheability. 176 00:20:11,106 --> 00:20:20,506 Through independent encoding of requests and options, Jeva ensures efficient and reusable representations, which streamline decision-making processes. 177 00:20:20,550 --> 00:20:36,990 Evan: Right, and the architecture's use of a compact decision stack and symmetric aggregation within decision blocks allows for efficient handling of varying candidate sets without compromising performance. 178 00:20:37,038 --> 00:20:40,668 Ashley: Jeva's performance in empirical evaluations is impressive. 179 00:20:40,668 --> 00:20:52,928 It excels in first-argmax accuracy, null detection, and permutation stability tests, maintaining high accuracy and robust decision-making across different permutations of candidate sets. 180 00:20:52,928 --> 00:21:00,798 Its ability to reduce latency significantly through caching mechanisms adds practical value for real-world applications. 181 00:21:00,846 --> 00:21:15,066 Evan: In essence, Jeva's architecture could transform how we approach rapid decision-making in conversational AI, tool selection, and document retrieval tasks, offering a scalable and efficient solution. 182 00:21:15,126 --> 00:21:16,126 Ashley: Indeed. 183 00:21:16,126 --> 00:21:33,126 By separating reusable semantic encoding from set-dependent judgment and introducing mechanisms for permutation invariance and null-awareness, Jeva stands out as a robust and innovative decision-making model that meets the demands of modern AI applications. 184 00:21:33,174 --> 00:21:46,094 Evan: And that brings us to the end of today's discussion on 'Jeva: System One Models from First Principles.' We hope you found this deep dive into their contributions as insightful as we did. 185 00:21:46,158 --> 00:21:48,018 Ashley: Thank you for joining us! 186 00:21:48,018 --> 00:21:56,978 Make sure to subscribe and tune in every day for more insights into the latest research in AI, NLP, CV, and beyond. 187 00:21:57,090 --> 00:22:02,110 Evan: If you have any questions or papers you'd like us to cover, feel free to reach out. 188 00:22:02,110 --> 00:22:04,890 We're always excited to hear from our listeners. 189 00:22:05,000 --> 00:22:10,580 Ashley: Until next time, stay curious and keep exploring the frontiers of AI. 190 00:22:10,580 --> 00:22:11,950 Have a great day! 191 00:22:12,006 --> 00:22:16,316 Evan: See you in the next episode of Daily Paper Cast. 192 00:22:16,316 --> 00:22:17,826 Bye for now!